Technology

Compute is now geopolitical. US GPU export controls whipsaw global chip supply and pricing; Japan is turning digital sovereignty into a procurement requirement via Rapidus 2nm and sovereign clouds; and India is rising as a new compute geography on the back of its semiconductor mission and mega data-center investment. On top of that, AI-agent interoperability protocols (MCP, A2A) are moving under neutral standards bodies (the Linux Foundation / AAIF). Where, on whose silicon, and on whose standard you compute is increasingly set by policy and governance.

1. 🇺🇸 US AI-chip policy is whipsawing: after the Jan 2025 'AI Diffusion Rule' and the April H20-to-China licensing requirement, December 2025 conditionally allowed NVIDIA H200 exports to China (capped at 50% of US-customer volume, a 25% tariff, codified into a formal BIS rule on Jan 13 2026), while in Congress a Chip Security Act bill (mandating location-tracking tech in chips) and a Senate bill proposing a 30-month ban on H200/Blackwell to China (the Secure and Feasible Exports Chips Act) are under debate — feeding directly into uncertainty over global GPU supply, pricing and procurement lead times.[21][31]
2. 🇯🇵 Japan is turning digital sovereignty into a procurement requirement: Cloudflare completed ISMAP (the government's security-assessment program) registration on Jan 15 2026, letting it serve public agencies with its Data Localization Suite and Geo Key Manager (SSL/TLS keys held in Japan); SoftBank is standing up a sovereign cloud on Oracle Alloy (East Japan Apr 2026, West Oct 2026) plus a sovereign AI GPU cloud; and the government — Rapidus's top shareholder (11.5% voting rights) with a golden share — is backing 2nm mass production (targeted 2027). Policy actively supports the option to 'compute domestically, on domestic silicon.'[23][24][26]
3. 🇮🇳 India is emerging as a new node for chips and compute. Under the India Semiconductor Mission (ISM 2.0), Micron's Sanand (Gujarat) ATMP plant came online in Feb 2026 (~$2.75B), Tata Electronics x Taiwan's PSMC are targeting first silicon at the Dholera fab in late 2026, and there are 13 approved projects with an ~₹80B allocation in the 2026-27 budget. As a hedge against US-China friction, an alternative geography for assembly/packaging and future logic production is standing up.[27][37]
4. Post-quantum cryptography is a government-driven migration: NIST's IR 8547 (Nov 2024) lays out a phase-out plan for quantum-vulnerable RSA/ECC, and US federal systems are to move fully to PQC by 2035 — making cryptography an area where standards bodies and state policy effectively set the timetable.[9]

Coming soon

AI inference cost fell an order of magnitude with the Blackwell generation, transforming the economics of shipping AI features; meanwhile game-engine pricing is being reshaped (Unity raising subscriptions) and AI dev tooling is shifting from flat fees to metered usage (GitHub Copilot). Agent payments (ACP / AP2) open a new transaction layer of purchases made via chat/agents. Compute geography is also diversifying — India is adding cheap data-center capacity fast while Japanese investment still faces power/cost pressure. Compute and licence costs need re-modeling.

1. AI inference cost is collapsing: per-million-token cost fell from ~20c on Hopper to ~10c on Blackwell and ~5c with low-precision NVFP4, with NVIDIA claiming ~$0.12 on the GB300 NVL72 (35x lower than Hopper); leading inference providers report 4-10x cost reductions, upending the economics of putting AI features into products.[14][15]
2. 🇮🇳 India is rising fast as a 'cheap compute geography': a large wave of investment (reportedly cumulative hundreds of billions of dollars) is flowing into India's AI ecosystem; Google broke ground on a Visakhapatnam (Andhra Pradesh) AI hub (~1GW; ~$15B over 2026-2030), Reliance is planning a world-leading 3GW data center in Jamnagar, with Meta the first built-to-suit tenant at 168MW, and capacity is set to hit ~7GW by 2030. At lower cost than other APAC and Western markets, it is becoming a realistic sourcing option for inference/training workloads.[28][29]
3. 🇯🇵 Japan's data-center market is set to grow from ~$12.8B in 2025 to ~$38.9B by 2031 (~20% CAGR), with Tokyo alone hosting ~74 existing plus 26 upcoming sites as of June 2025; supply and competition rise, but AI-driven power constraints remain an upside risk to the cost base.[16]
4. Game-engine cost structures are being reshaped: Unity scrapped the install-based Runtime Fee but raised subscriptions (Pro +8% to $2,200/seat in Jan 2025, Enterprise +25%, with further increases and an extended free tier in Nov 2025). Alongside royalty-free Godot and Unreal's royalty model, this forces a recalculation of build-vs-license TCO.[4][5][6]
5. 🇺🇸 AI dev tooling is shifting from flat fees to metered usage: GitHub Copilot moved to usage-based billing (GitHub AI Credits, per-token) on June 1 2026 — completion stays free but Chat, CLI and cloud agents consume credits (Pro $10 with $15 credits, Pro+ $39 with $70, Max $100 with $200). AI-coding spend becomes a variable cost like infrastructure, forcing a redesign of seat-based tooling budgets and FinOps.[36][18]
6. AI coding productivity is now a cost-benefit question: DX analysis finds ~3.6 hours/week saved per developer and Copilot is used by ~90% of the Fortune 100, but GitClear flags quality erosion such as rising code duplication (clones) — so gains must be weighed against review and maintenance cost.[18][19]
7. Agent payments are becoming a new transaction layer: in Sept 2025 OpenAI x Stripe launched Instant Checkout inside ChatGPT and the 'Agentic Commerce Protocol (ACP)', and on Sept 16 Google announced the 'Agent Payments Protocol (AP2)' with 60+ partners (Mastercard, PayPal, American Express, Coinbase and more). AP2 governs agent purchases via cryptographically signed 'Mandates', supports stablecoins/crypto through the A2A x402 extension, and composes with MCP/A2A. The funnel is shifting from app to chat/agent-mediated purchase, reshaping how payment and conversion economics are split.[43][44]

Coming soon

AI-assisted development has become standard yet trust is falling, prompting a rethink of hiring, training and review culture; India is maturing as the world's largest AI/engineering talent pool for sourcing, while free, offline on-device AI is normalizing, making local processing of photos and children's data a new privacy expectation among users. On top of that, agentic browsers are rewiring discovery, traffic and purchase paths, making a service's 'agent readability' a new acquisition channel.

1. Stack Overflow's 2025 survey shows 84% use or plan to use AI tools (up from 76%) and 51% of professionals use them daily, yet positive sentiment fell to 60% (from 70%+ in 2023-24), only 3.1% 'highly trust' the output and 45.7% distrust it — shifting the premises of hiring criteria and review processes.[18]
2. 🇮🇳 India is the world's largest engineering and AI talent pool: Global Capability Centers (GCCs) number 1,800+ employing ~2M people, with Bengaluru alone hosting 880+ units and ~660k professionals (~36% of the total), and Fortune-500 India GCCs holding ~126,600 AI-aligned roles. Placing software / AI-ML / cloud / security capability in a cost-efficient offshore geography has matured into a real option.[33]
3. On-device AI is raising privacy expectations: Apple's Foundation Models (free, offline, no API key) and Android's Gemini Nano / ML Kit GenAI are spreading, making local processing of sensitive data like photos and children a baseline user expectation — a tailwind and an accountability question for a family-data product like FamilyAlbum.[1][2]
4. Passwordless authentication is going mainstream: per the FIDO Alliance, over 1 billion people have activated at least one passkey and 15B+ accounts support them, while Google reports 800M accounts and 2.5B sign-ins with ~30% higher success rates. Passkeys cut account takeover, support cost and drop-off — bearing directly on login for the payments-handling PointsBet and on mixi2/FamilyAlbum.[10]
5. Agentic browsers are rewiring discovery and traffic: Perplexity's Comet (July 2025) and OpenAI's ChatGPT Atlas (Oct 2025) have surged, and browser-based agents accounted for ~71% of agentic traffic in April 2026, with Comet ~47% and Atlas ~20.3% in May 2026. Because AI browses, summarizes and buys on the user's behalf, the assumptions of web traffic, referrals and SEO break down, and making content/services structured and API-accessible — their 'agent readability' — becomes a new acquisition channel.[46][47]
6. AI-native development is redefining engineering culture: terminal-first Claude Code (defaulting to the latest Claude Opus 4.8, state-of-the-art on SWE-bench Verified, up to 1M-token context on Max) and IDE-first Cursor are the two dominant agentic dev environments, and many production teams use both — line-level edits plus goal-level delegation. The role shifts from 'writing code' to 'delegating goals and supervising autonomous execution', putting design, review and acceptance-criteria skills at the center of hiring and training.[45][18]

Coming soon

On-device AI going mainstream (Apple Foundation Models, Android Gemini Nano), a game-engine generational shift (Unity 6.2 with Unity AI, the Unreal Engine 6 reveal), WebGPU reaching Baseline, and data-platform standardization on Apache Iceberg (v3 GA, v4 in progress) are advancing together. Layered on top, agent infrastructure (MCP, A2A), AI-native dev stacks (Claude Code / Cursor) and edge/post-cloud inference (Cloudflare Workers AI) are renewing the application layer — rewriting how the client, server, data and agent layers are built.

1. On-device AI has graduated into a product API. At WWDC 2026 Apple overhauled Foundation Models with a multimodal on-device model that handles images (8,192-token context), plus Private Cloud Compute (32K tokens, configurable reasoning, free for apps under 2M first-time downloads) and a `LanguageModel` abstraction layer that can swap in Anthropic, Google or local MLX models; Android similarly offers Gemini Nano on AICore (ML Kit GenAI: summarize, rewrite, image description, with a Prompt API in alpha).[1][2][3]
2. Game engines are at an inflection point: Unity scrapped the Runtime Fee (Sept 2024) and shipped Unity AI and automatic LOD generation in Unity 6.2 (Aug 2025). On mobile, Unity holds ~48% share and powers ~70% of top-grossing titles, even as Epic revealed Unreal Engine 6 in May 2026 (Early Access targeted for late 2027) and royalty-free Godot rises — putting medium-term client-tech choices on the table.[4][5][7]
3. Web standards leveled up: WebGPU reached 'Baseline' in Jan 2026, shipping stably across every major browser — Chrome, Edge, Firefox and Safari 26+ — enabling near-native GPU rendering and compute in the browser. WebNN remains an origin trial (not production) on Chromium, with WebAssembly (3.0) as the universal fallback; together they make rich off-app-store (web-distributed) games and on-device inference materially more viable.[8][34][22]
4. Data platforms have decisively converged on open table formats: Apache Iceberg is the de facto standard, and at the 2026 Iceberg Summit (600+ attendees) Snowflake, Databricks, AWS, Google and Microsoft all read and write it. Both vendors made Iceberg v3 GA (Snowflake via Horizon/Open Catalog, Databricks via Unity Catalog), the REST Catalog has become the lakehouse integration hub, and a v4 spec is in progress — standardizing designs that decouple storage from query engine and avoid vendor lock-in.[17][35]
5. Agent infrastructure is standardizing fast. Anthropic's Model Context Protocol (MCP) hit ~97M monthly SDK downloads by March 2026 (a ~970x rise in 18 months) with 10,000+ public MCP servers, and its Nov 2025 spec added task-based workflows, authorization extensions and URL-based client registration. Together with Google's A2A (Agent Card-based discovery and collaboration between agents, 150+ organizations), the 'plumbing' that connects AI agents to external tools and other agents is solidifying — making whether FamilyAlbum / mixi2 / Monster Strike / PointsBet are 'callable by agents' a real tech-choice question.[41][39][42]
6. AI-native dev stacks have taken hold: terminal-first Claude Code (defaulting to the latest Claude Opus 4.8, state-of-the-art on SWE-bench Verified, up to 1M-token context on Max) and IDE-first Cursor (sub-second tab autocomplete, background cloud agents) are the two dominant agentic dev environments, with OpenAI Codex and Cline following. Development is shifting from code generation toward agentic orchestration — delegating goals and letting agents execute autonomously.[45]
7. Inference is distributing to the edge / post-cloud. Cloudflare Workers AI runs serverless inference of models like Llama 4 Scout, Mistral and Kimi across 330 cities and 300+ edge locations, delivering sub-100ms latency to most users, and ships new protocols such as Pay Per Crawl (July 2025, charging crawlers via HTTP 402). Computing near the user rather than in a single central data center is becoming real, widening the design room for latency- and data-locality-sensitive workloads (odds, recommendation, image processing).[48]
8. 🇯🇵 This bears directly on MIXI's product stack: Monster Strike is built on Unity and exposed to engine-generation/pricing shifts, FamilyAlbum runs image ML and auto-editing well suited to on-device inference, and the PointsBet consolidation brought in-house real-time infrastructure for payments, KYC and low-latency odds. On-device AI, data-platform standardization, crypto migration and sovereign cloud (SoftBank x Oracle / ISMAP) — plus agent infrastructure (MCP/A2A), agent payments and edge inference — all act on these three products' tech choices at once.[1][17][24]
9. AI-agent interoperability is moving under neutral standards bodies. Anthropic donated the Model Context Protocol (MCP) to the newly formed Agentic AI Foundation (AAIF) under the Linux Foundation on Dec 9 2025 (co-founded by Anthropic, Block and OpenAI, backed by Google, Microsoft, AWS, Cloudflare and Bloomberg), while Google's Agent2Agent (A2A) was transferred to the Linux Foundation in June 2025 with 150+ organizational supporters. The 'plumbing' of the agent economy is being standardized under open governance rather than one-vendor control, making it easier to avoid lock-in.[39][40][42]
10. The open-source supply chain has become an attack surface: the self-replicating 'Shai-Hulud' worm of Sept 2025 poisoned 500+ npm packages (prompting a CISA alert), and December's 'Shai-Hulud 2.0' spread to 25,000+ repositories and into Maven/PyPI. With credential and secret theft now routine, dependency pinning, least-privilege tokens and signed provenance are becoming mandatory.[11][12]

Coming soon

Security and data protection have become matters of law and compliance: the EU CRA imposes vulnerability reporting and security requirements on 'products with digital elements', NIST's post-quantum migration sets 2030 deprecation and 2035 disallowance deadlines, India's DPDP, Australia's Privacy Act reform, Japan's (promotion-oriented) AI Act plus per-country data residency fragment data and AI governance, and a wave of npm supply-chain worms is making SBOMs and signed provenance de facto requirements.

1. The EU Cyber Resilience Act (CRA) is entering its enforcement phase: it imposes security requirements on 'products with digital elements' placed on the EU market, with mandatory reporting of exploited vulnerabilities and severe incidents (24-hour early warning, 72-hour notification) starting Sept 11, 2026 and full application on Dec 11, 2027 — broadly covering software such as apps.[13]
2. 🇮🇳 India's DPDP is entering real enforcement: MeitY notified the Digital Personal Data Protection Rules 2025 on Nov 14 2025 and the Data Protection Board of India was instituted on Nov 13. Operational duties on consent, notice, breach notification, cross-border transfer and children's data phase in over 18 months — full compliance around May 2027, consent-manager registration from Nov 2026. Any design that serves Indian users or processes personal data via an Indian GCC will need to comply.[30]
3. 🇦🇺 Australia's Privacy Act reform has moved into an operational phase: the OAIC ran its first compliance sweep in Jan 2026 (~60 organizations across six sectors) and is wielding enhanced powers (civil penalties up to A$50m for serious/repeated breaches, compliance notices, breach-investigation rights). From Dec 10 2026, automated decision-making (ADM) transparency duties take effect, requiring ADM disclosures in privacy policies — bearing directly on PointsBet, which runs payments, KYC and odds automation in the AU market.[38]
4. 🇯🇵 Japan has enacted a promotion-oriented (innovation-first) AI Act: the 'Act on Promotion of Research, Development and Utilization of AI-related Technologies' passed on May 28 2025 and came into full effect on Sept 1 2025. It is a penalty-free basic law — in contrast to the prohibition-heavy EU AI Act it leans toward promotion (state duties, basic policy, operator cooperation). Interpretive guidance on existing laws (Copyright Act, APPI) and the Cabinet Office's work on IP-protection/transparency principles for generative AI remain live operational questions. It is a tailwind for AI deployment in Japan, but data and copyright handling must be built into designs.[49][50]
5. Post-quantum migration now has deadlines: NIST IR 8547 sets quantum-vulnerable RSA-2048 and ECC P-256 to be deprecated by 2030 and disallowed by 2035, with general encryption moving to ML-KEM and signatures to ML-DSA/SLH-DSA. Long-retained data faces a 'harvest-now, decrypt-later' threat, bearing on the crypto design of family photos and payment data.[9]

Coming soon

AI data-center power demand is surging — in the US the grid has become the binding constraint on AI infrastructure — pushing up the cost and sustainability of cloud compute; conversely, the shift to on-device inference (Apple AFM, Gemini Nano), edge inference (Cloudflare Workers AI) and Blackwell-generation efficiency act as levers that cut per-task energy and cloud load.

1. The IEA projects data-center electricity use will more than double to about 945 TWh by 2030 (roughly Japan's total consumption); AI-concentrated demand puts upward pressure on both cloud-compute cost and ESG/sustainability, flowing through to infra costs for storage/distribution-heavy services like MIXI's.[20]
2. 🇺🇸 In the US, power has become the binding constraint on AI infrastructure: data-center demand is projected (per one forecast; estimates vary widely, e.g. 31->41->66GW elsewhere) to grow from ~62GW (2025) to ~76GW (2026) to ~134GW (2030), and PJM's 2027/28 capacity auction fell ~6GW (6,625MW) short of its reliability target for the first time; component shortages (transformers, switchgear) and local opposition mean only about a third of 240GW planned is actually being built, and hyperscaler capex growth decelerates in 2026. OpenAI's Stargate (~7GW planned, $400B+) has stalled in Texas — feeding directly into cloud-compute supply, price and siting risk.[32]
3. On-device/edge AI and inference efficiency are mitigation levers: on-device models that move inference to the device/NPU (Apple Foundation Models, Gemini Nano), edge inference across 330 cities (Cloudflare Workers AI) and the order-of-magnitude inference efficiency gains on Blackwell cut per-task energy and cloud load — widening the design room to align cost optimization with lower environmental impact.[1][14][48]

Coming soon